Speaker
Description
Large three-dimensional k-Wave simulations are often limited by the memory capacity of a single GPU, while CPU execution can be prohibitively slow. We present a feature-complete multi-GPU CUDA solver that preserves the global k-space operator, supported physical models, and existing k-Wave workflow while distributing simulation fields across multiple devices. Global three-dimensional FFTs use cuFFT Xt to redistribute slab-partitioned data between spatial and spectral layouts. CPU, single-GPU, and multi-GPU execution are integrated within one unified C++17 and CUDA codebase with shared input, output, logging, and regression infrastructure. The solver is continuously validated against 137 MATLAB-reference simulations, performing 859 output-dataset comparisons per execution mode across multiple HPC systems. On Karolina, eight A100 GPUs achieve up to 29.3 times speedup over the optimized 128-thread OpenMP solver. Distributed GPU memory also enables a measured $1400^3$ simulation containing 2.744 billion grid points.